on 29-Nov-2024 (Fri)

The learning mechanism of the recurrent neural network thus involves:

...

(2) the backpropagation step where the gradient of the parameters with respect to the loss is calculated; and finally,

...

Adding to the complexity of ML projects are two other critical elements that are not shared by (most) traditional software development projects: a frequent lack of detail in project expectations and the relative industry immaturity in tooling.

Using an RNN where all firm actions and customer responses are represented by elements in a vector. For instance, suppose a firm solicits customers either through phone, mail, or email (three channels), and customers may purchase across 17 product categories. All the analyst has to do is to encode each observation period (e.g., a day, a week, a month) as a vector of size 20, where all the values are equal to 0, except when a solicitation is sent, or a purchase is observed.

Coronaviruses are enveloped positive-sensed, single-stranded RNA viruses.

ЅARЅ-СoV-2 is a betacoronavirus. It is in the same subgenus as, but is in a different clade from, the severe acute respiratory syndrome (ЅARЅ) virus

The Middle East Respiratory Syndrome (MERS) virus is a more distantly related betacoronavirus [ 7,8].

The closest RNA sequence similarities to ЅARЅ-CoV-2 are found in coronaviruses isolated from bats, which were the likely animal reservoir prior to spillover to humans [ 9,10]. Pangolins are a possible intermediate host for ЅΑRЅ-СoV-2 prior to passage to humans.

The primary host cell entry receptor for ЅΑRЅ-CоV-2 is the same as for SARS-CoV, the angiotensin-converting enzyme 2 (ACE2) [11]. ЅΑRS-СоV-2 binds to ACE2 through the receptor-binding domain of its spike protein (figure 1).

Additional enzymatic interactions between ЅΑRS-CoV-2 and host cells determine specific cell tropism, which has evolved with the virus. Specifically, following ACE2 binding, the original ЅARЅ-СoV-2 virus and some subsequent variants, including Delta, enter cells after a host protease (most often the TMPRSS2 enzyme) cleaves the spike protein, enabling subsequent membrane fusion and cell entry [12,13]. Since TMPRSS2 is expressed at high levels in lung epithelial cells, these variants replicated efficiently in the lungs. Omicron-lineage variants, in contrast, mostly enter host cells through a different pathway involving cathepsin-dependent endocytosis and use the TMPRSS2-mediated pathway far less efficiently [14]. Thus, Omicron-lineage variants replicate more efficiently in the upper airways, where TMPRSS2 expression is lower than in the lungs [15].

ЅΑRS-CoV-2 antivirals have targeted one of two important viral proteins:

● RNA-dependent RNA polymerase (RdRp) – This is an error-prone enzyme responsible for viral replication [16]. Remdesivir, molnupiravir, and mindeudesivir (a drug available in Сhiոa) target the RdRp.

● 3CL protease – This is also called the main protease, and it cleaves the polyproteins encoded by the ЅΑRЅ-СоV-2 RNA. Ritonavir-boosted nirmatrelvir [17], ritonavir-boosted simnotrelvir, and ensitrelvir target this protease.

ЅΑRS-CoV-2, like many other RNA viruses, has evolved over time through random mutations from replication infidelity, although its intrinsic mutation rate is lower than many other RNA viruses (including HIV and HCV) because of a proofreading mechanism [18].

Recombination is another mechanism for evolution. This occurs when a host is simultaneously infected with two distinct ЅARЅ-СоV-2 viruses, and a progeny virus emerges that includes RNA sequences from both [19].

Rapid accumulation of many mutations may occur in highly immunocompromised individuals with persistent or chronic ЅARЅ-СoV-2 infections [18,20,21]. (See 'Viral shedding and period of infectiousness' below.)

Most importantly, selective advantage to a progeny virus mainly comes from intrinsic viral factors (often leading to replication advantage) or from greater evasion of the host immune response (most commonly from antigenic shift)

The World Health Organization (WHO) has designated labels for notable variants based on the Greek alphabet [22]. Each variant also has several designations based on the nomenclature used by distinct phylogenetic classification systems.

The Pango dynamic nomenclature is a commonly referenced and important system, in which each major lineage begins with a letter, followed by additional letters or numbers (eg, Β.1.1.529 or BA.1 are the Pango delineations of the original Omicron virus); recombinant viruses include an “X” in their Pango nomenclature [ 23,24].

Since the emergence of ЅΑRЅ-CоV-2, several mutations have been shown to convey an intrinsic replication advantage independent of immune escape mechanisms:

• The D614G spike mutation enhances binding of the virus with the host cell ACE2 receptor, leading to a 20 percent increase in transmissibility [25].

• The N501Y substitution also conveys an inherent replication advantage over wild-type viruses [26]. This substitution is associated in increased affinity between the spike protein and ACE2; virus containing this mutation achieves a higher effective viral load.

Until around September 2020, there was remarkable stability among ЅARЅ-СoV-2 virus genomes, as the population was essentially fully susceptible and there was little selective pressure from population-level immunity to drive viral evolution.

The Alpha variant (Β.1.1.7 lineage) was approximately 50 to 75 percent more transmissible than prior ЅΑRS-СοV-2 viruses and became globally dominant in late 2020 until the emergence of the Delta variant [27-33].

The Beta (Β.1.351 lineage) and Gamma (P.1 lineage) variants became regionally dominant in late 2020 in South Africa and Brazil, respectively, but were not globally dominant [34,35].

Delta emerged in December 2020, became dominant worldwide in mid-2021, and remained so until emergence of the Omicron variant [41,42]. Delta was more transmissible than any of the prior variants.

The Omicron variant, with around 50 novel mutations compared with wild-type virus, including 34 in the spike gene, was first reported from Botswana and very soon thereafter from South Africa in November 2021 [43]. It rapidly spread around the globe, replacing all previously dominant circulating ЅARЅ-CoV-2 viruses. Subsequently, Omicron sublineages have repeatedly emerged and replaced the previous predominant sublineage; each prevailing sublineage has some advantage over the prior one, whether it is increased intrinsic transmissibility or increased evasion of existing immune responses or both (table 1).

Omicron sublineages have a number of features that distinguish them from previously circulating variants

Omicron sublineages are more transmissible compared with prior ЅΑRЅ-CoV-2 variants. Intrinsic transmissibility of a virus can be estimated by the basic reproductive number (R0), which is the number of secondary cases caused by each index case among a fully susceptible population. The R0 for ЅΑRЅ-CоV-2 increased from 2.5 for the original virus [44] to around 5 for Delta [45] and to greater than 8 for Omicron [46]

The tropism of the BA.1 virus for the nasopharynx may have contributed to increased transmissibility compared with prior ЅΑRS-СοV-2 viruses, since it could more readily establish an infection in the more accessible nasopharynx and upper airways [18].

Omicron more effectively evades pre-existing immunity than prior ЅΑRЅ-CοV-2 viruses [49,50]. In a case-control study from Qatar, a history of prior infection was associated with an 85 to 90 percent lower risk of infection with Alpha, Delta, or Beta variants, but only a 56 percent lower risk with Omicron BA.1 [50].

Similarly, compared with Omicron BA.1, sublineages BA.4 and BA.5 were not as well recognized by antibodies elicited by BA.1 or BA.2 infection or vaccination [54-56]

However, prior infection with one Omicron subvariant may still provide some protection against subsequent infection with certain other subvariants, even if in vitro studies suggest low antibody cross-reactivity [ 57-59]. As an example, the risk of reinfection with BA.4 and BA.5 was lower following BA.1 or BA.2 infection than following infection with a pre-Omicron variant [57,58]

Omicron sublineages have also escaped neutralization by monoclonal antibodies derived from individuals infected with prior ЅARЅ-CoV-2 viruses, thus limiting the preventive and therapeutic benefits of such interventions.

T-cell immunity against different ЅΑRS CoV-2 variants, including Omicron, is more preserved, likely explaining protection against severe outcomes for individuals with prior infection and/or vaccination [60]

Unlike antibody responses, which are focused on immunodominant regions of the viral spike protein that are prone to escape from humoral immunity, T-cell responses after infection and vaccination are broad and directed against conserved viral epitopes.

Evasion of innate immune responses – This is another important mechanism for ЅARЅ-СοV-2 immunopathogenesis. For example, the Omicron subvariants BA.4 and BA.5 had higher levels of Orf6 gene expression leading to antagonism of host innate immunity [62]

However, certain symptoms may have been more prominent with specific variants, in part because of different tropism. As an example, sore throat has been more prevalent with Omicron sublineage infections, while loss of smell or taste was more prevalent with Delta and pre-Delta infections [63]

Observational data from multiple studies suggest that the risk of severe disease or death with Omicron infection is lower than with prior variants of concern [64-73]. An analysis from England estimated that the risk of hospital admission or death with Omicron was approximately one-third that with Delta, adjusted for age, sex, vaccination status, and prior infection [67]

The reduced risk for severe disease may reflect partial protection conferred by prior infection or vaccination. However, animal studies show lower viral levels in lung tissue and milder clinical features (eg, less weight loss) with Omicron compared with other variants; this provides further support that Omicron infection may be intrinsically less severe [74-76].

However, the magnitude of antibody response may be associated with severity of disease, and patients with mild infection may not mount detectable neutralizing antibodies [79,80]

Following infection with ЅARS-СοV-2, most patients develop detectable serum antibodies to the receptor-binding domain of the viral spike protein and associated neutralizing activity [77,78]

Cell-mediated immunity – Studies have also identified ЅΑRЅ-CоV-2-specific CD4 and CD8 T-cell responses in patients who had recovered from СΟVІD-19 and in individuals who had received СОVID-19 vaccination, which suggest the potential for a durable T-cell immune response [81,88,94,95]

Q. What type of chicken parts are used in stock?

A. Bones.

This question focuses on just one detail. It’s precise about what it wants (“What type of chicken parts…?”), and it should produce a consistent answer over time. It’s plenty tractable, but it still demands the effort of retrieving the answer from memory.

Writing a simple factual prompt like that naturally tickles a neighbor you might consider adding: the explanation prompt. I write prompts like this when a detail seems likely to be challenging or when the explanation itself is interesting. A more experienced cook likely wouldn’t bother with the first question, but they might still find this one useful.

Q. Why do we use bones to make chicken stock?

A. They’re full of gelatin, which produces a rich texture.

Note that the answer is found in a different section of the text! Writing prompts often involves hopping around as you work to identify the puzzle pieces and put them together.

Now, this explanation prompt is a good first try, but it could be more precise. It would be reasonable to answer “Why do we use bones to make chicken stock?” with “Because bones are economical.” You want to write questions which cause you to unambiguously retrieve the information you have in mind. This more precise prompt is better:

Q. How do bones produce a chicken stock’s rich texture?

A. They’re full of gelatin.

From here, we might ask:

Q. What are typical aromatics used in chicken stock?

A. onion, carrots, celery, garlic, and parsley

But unless you’re an experienced cook, you’ll probably find this prompt intractable, or you may remember ingredients inconsistently. Unordered lists like this can be challenging to translate into good prompts.

One good strategy is to create a set of questions which require you to fill in a missing element of the list:

Q. Typical chicken stock aromatics:

• ???
• carrots
• celery
• garlic
• parsley

A. Onion

Q. Typical chicken stock aromatics:

• onion
• ???
• celery
• garlic
• parsley

A. Carrots

… and so on. Note that I keep the list in the same order. When each element has a consistent (though arbitrary) location, you’ll end up learning the list’s visual “shape” to some extent as you repeat these prompts. That will help your recall.

Such cues engage another memory phenomenon cognitive scientists have explored experimentally: you make information easier to recall when you connect it to other memories. This process is called elaborative encoding See e.g. Bradshaw and Anderson, Elaborative Encoding as an Explanation of Levels of Processing (1982). . The members of an ingredient list can be difficult to relate to anything meaningful. In such cases, you can still leverage elaborative encoding by fabricating an association as a mnemonic device. Vivid associations work best, so it’s helpful to find relationships involving visuals, meaningful personal experiences, or emotions like humor and disgust.

Putting these mnemonic devices in the answer field is a useful general trick when the information you’re trying to remember seems relatively arbitrary or isolated.

Instead of fabricating a mnemonic device, you can also leverage elaborative encoding by adding imagery to your prompts. The same fill-in-the-blank idea applies quite well to images. You’d probably find this prompt much easier to remember over time than its textual equivalent:

Q. Chicken stock ingredients:

A. Carrots

Of course, cues and elaborative encoding add extra effort. You don’t need to use them all the time: you’d probably find that exhausting. But they’re useful techniques to deploy if you suspect a prompt will prove difficult to remember, or if you’ve noticed yourself struggling in practice

Writing good prompts often involves interpretation, a first step to creating understanding beyond what’s explicitly written.

Q. What’s the ratio of chicken bones to water in chicken stock?

A. A quart of water per pound of bones

Q. How much onion to use in chicken stock per pound of chicken bones?

A. Half an onion per pound

Q. How much carrots/celery to use in chicken stock per pound of chicken bones?

A. 1 carrot/celery rib per pound

Q. How much garlic to use in chicken stock per pound of chicken bones?

A. 2 smashed cloves per pound

Notice how I’ve broken the ingredient list down into many questions here, each focused and precise. I’ve noticed that people often feel a compulsion to economize on the number of prompts they write. Prompts seem to carry a per-unit “price,” so people naturally try to write fewer questions which cover more ground. But that’s counter-productive. Unless you explicitly decide to exclude certain information, the number of “units of raw knowledge” is fixed, a constant of the territory. When you write coarser prompts in smaller quantity, you’re not reducing the amount you have to learn. You’re just making the material harder to review.

Prompts are cheaper than you probably imagine. An easy prompt will consume 10–30 seconds across the entire first year of practice, and much less in each subsequent year. Until you’ve internalized that observation, try to adopt this rule of thumb: write more prompts than feels natural.

Now, prompts are cheap, but they’re not free. Besides their time cost, they have an emotional cost: no one wants to spend time reviewing a bunch of boring material they already know. So if you’re writing prompts for a subject that’s already quite familiar, you should use fewer prompts—not because it’s always safe to write coarser questions for familiar topics, but because there’s less marginal knowledge you need to capture.

Relatedly, the appropriate scale of a “focused” prompt depends on the scale of the concepts you’ve internalized. This particular set of aromatics is so deeply familiar as a group that I’d write prompts which treat it as a unit (“Italian aromatics”) instead of memorizing individual ingredients and quantities.

As you build fluency in increasingly complex concepts, you can write increasingly complex prompts while keeping each focused on what feels like a single detail. In fact, the ability to think in terms of increasingly complex “chunks” appears to be a significant part of what expertise actually is. For a compelling demonstration, see Chase and Simon, Perception in chess (1973), which experimentally demonstrates how chess masters operate in terms of larger chunks. Viewed through this lens, one role for memory systems is to accelerate the process of increasing your effective chunk size in a topic.

Procedural knowledge

Now that we’ve got prompts for the ingredients, let’s look at the steps of the recipe. This is procedural knowledge—knowledge needed to perform specific tasks, more knowing how than knowing what.

In some sense, procedures are lists. So we can start by using the cloze-deletion method we used for the ingredient list:

Q:

1. ???
2. Bring to a simmer on low heat (this will take about an hour). We use low heat to produce a bright, clean flavor: at higher temperatures, the stock will both taste and look duller.
3. Lower heat to maintain a bare simmer for an hour and a half.
4. Strain, wait until cool, then transfer to storage containers.

A. Combine all the ingredients in a large pot.

… and so on, for each of the four steps.

This is an awfully unfocused prompt. It includes many unimportant details which distract from the knowledge you actually intend to retrieve. In my experience, wordy prompts like these tend to dull my concentration and produce vague, distracted answers.

We can improve this prompt somewhat by simply removing words. If necessary, we can add extra prompts to cover any details we removed.

Q:

1. ???
2. Slowly bring to a simmer
3. Maintain bare simmer for 90m
4. Strain, wait until cool, then store

A. Combine all ingredients

This is better. But editing down the procedure helps us make a few observations about this knowledge. First: a few keywords (or word groups) carry the critical details of the procedure: slowly bring to a simmer, then maintain bare simmer for 90m. If you know those four bolded phrases, you basically know this procedure. The other words are just a skeleton. Which brings us to a second observation: the first and fourth steps aren’t worth writing prompts about. If you know what stock is (and we’ll get to that), you know you’re simmering a bunch of ingredients in a pot. When the stock is finished simmering, the final step is common sense. What else would you do?

After highlighting keywords in this way, prompt-writing often feels like playing Jeopardy: can you phrase that in the form of a question? Here’s a revised set of questions which capture my understanding of this procedure.

Q. At what speed should you heat a pot of ingredients for chicken stock?

A. Slowly.

Q. When making chicken stock, when should you lower the heat?

A. After the pot reaches a simmer.

Q. When making chicken stock, what should you do after the pot reaches a simmer?

A. Lower the temperature to a bare simmer.

Q. How long must chicken stock simmer?

A. 90m.

Procedures can often be broken down into keywords like this. What are the important verbs, and when should you move between them? What are the key adjectives, adverbs, subjects, objects?

In our stock recipe, the verbs aren’t very important: “bring,” “lower,” “strain.” You’re cooking ingredients in water at various temperatures, so those actions are obvious. But conditions or heuristics describing when to move between verbs are important: first when the water reaches a simmer, then after ninety minutes passes. And while it’s not worth capturing the fact that you’re heating the pot, the adverb “slowly” is indeed important.

The recipe is quite linear, but more complex procedures may branch, including conditions or special cases which could trigger some alternate path in the flowchart. Such predicate structures are usually worth capturing. If the branching is sufficiently complex, you might consider drawing a flowchart and using that in your prompt.

Is the keyword-based approach better than the list-based approach? That depends on how important the discrete details are and how intuitive they are for you. The keyword-based approach emphasizes the individual knowledge components more strongly, which is a good idea if precision is important. Also, I usually find focused questions like these more pleasant to answer. But if an outline is all you need, then the list approach is probably simpler.

It’s worth noting that in our editing, we’ve left behind a few useful details from the original recipe. Let’s re-add them with separate prompts:

Q. How long should it take to heat a batch of chicken stock (with 2lbs of bones)?

A. About an hour

This knowledge isn’t essential, but “heads-up!”-type information can be quite useful when learning procedures. If you know this detail, you’ll leave yourself enough time when making stock, you won’t be confused when your pot takes forever to heat up, and you might notice that you’re using too high a temperature if the pot boils in fifteen minutes.

Q. Why does Andy’s recipe claim we should prepare chicken stock over low heat?

A. “Brighter, cleaner” flavor.

Explanation-type prompts are especially valuable when studying procedures: they can help you avoid rote learning and build a deeper understanding. Note that in this particular case, we were only able to traverse one level of “why.” Why low heat? A cleaner flavor, says the recipe. Why does low heat produce a cleaner flavor? It doesn’t say. You might find yourself wondering what “brighter” or “cleaner” really even means. Writing this prompt revealed a gap in our understanding: how useful! We could dig into this question, but we’ll choose to keep moving for the moment. We indicate that by explicitly phrasing the answer as tentative, reliant on an external claim.

It’s helpful to add similar caveats to your prompts whenever you’re working with subjective, provisional, or incomplete information. In some sense, most of the prompts we’ve been writing are provisional: other recipes will probably do things somewhat differently. So we could phrase everything more tentatively, but in practice that’s more distracting than it’s worth. It’s usually enough to record where the information came from. Most spaced repetition systems have metadata fields you could use to link these prompts to the original chicken stock recipe.

What stock is and why it matters: conceptual knowledge

A wheel that can be turned though nothing else moves with it, is not a part of the mechanism.

— Ludwig Wittgenstein, Philosophical Investigations

Now we turn our attention to the first few paragraphs of that recipe. Definitions seem like the simplest things to write prompts for: after all, isn’t that how flashcards are most commonly used in school?

For instance, what about a pair of definition prompts, like this?:

Q. What is stock?

A. A flavorful liquid building block.

Q. Culinary term for flavorful liquid building block?

A. Stock.

Leaving aside quibbles about focus, precision, and consistency, we could probably memorize the verbatim answer to the first question. The reverse definition can help you remember the name for this term, which is often useful. But the ability to parrot these answers isn’t at all the same as knowing what stock is.

To internalize a concept, you need to understand its components and relationships. Your goal is to design a set of prompts which collectively trace the edges of “stock.”

I’ll now introduce some useful lenses for understanding concepts. You won’t necessarily need to use every lens for each concept you encounter. Think of them as a toolkit for identifying elements which seem most important to you. Most of these example prompts are best suited to a novice cook. Experienced cooks usually already know what stock is, though they don’t necessarily know how to make it.

Attributes and tendencies: What makes stock, stock? What’s always, sometimes, and never true of stock?

Q. How are stocks usually made?

A. Simmering flavorful ingredients in water.

Q. Why don’t stocks usually have a distinctive flavor?

A. To make them more versatile.

Similarities and differences: Knowing what stock is requires knowing what relates and distinguishes it from other adjacent concepts.

Q. How is stock different from soup broth?

A. Soup broth has a more complete flavor; stock isn’t meant to stand on its own.

Parts and wholes: What are some examples of stocks? Are there important “sub-concepts” of stocks? Is “stock” a part of some broader category? Visualize a Venn diagram, even if the edges are fuzzy.

Q. Name at least three examples of stock:

A. e.g. chicken, vegetable, mushroom, pork

Q. Stock is rarely served directly; it’s best thought of as a ??? (construction metaphor)

A. Building block.

Causes and effects: What does stock do? What causes it to do that? What doesn’t it do? When is it used?

Q. Why do restaurants use stock as a cooking medium instead of water? (name two reasons)

A. Adds flavor, improves texture.

Q. Stocks are a common base for… (name at least two)

A. e.g. sauces, soups, braises

Q. Restaurants often use stock as a cooking medium where home cooks might use ???.

A. Water

Significance and implications: Why does stock matter? What does it suggest? Make the concept personally meaningful.

Q. What liquid building block explains why simple restaurant dishes are often tastier than home renditions?

A. Stock.

Q. What should I ask myself if I notice I’m using water in savory cooking?

A. “Should I use stock instead?”

This last prompt isn’t stated in the recipe: it’s an example of what might be an understanding of your own. Notice, too, that its aim is more behavioral than intellectual. We’ll have more to say about using prompts to change behavior in our section on how to use chicken stock.