Week 1. Jan 10: Deep Learning? - Orienting

[avioberoi]

Post your question here about the orienting readings:

“Preface: How to Think with Deep Learning”, “Deep Learning?”, Thinking with Deep Learning, Preface & Chapters 1 & 2.

“On the Opportunities and Risks of Foundation Models.” 2022. Bommasani et al.

[youjiazhou]

For the "Deep Learning?" piece: How to think about the "weights" in the deep learning models? Why we assign weights to links instead of the sub-models themselves? How to think about the interrelationships among sub-models and how they can be compared to the structural equation model in traditional statistics?

[yangyuwang]

My question is about the number of neurons in the hidden layers.
Through most of the examples in "Deep Learning?" from Thinking with Deep Learning, it seems that the number of neurons in each layers are larger than the unit number in the input layer. (For example in Fig. 1-9, there are 2 input units and 3 neurons in each layer.) However, there are also transformers such as "superposition" (leveraging one neuron to represent multiple features), where the number of neurons should be less than the input unit numbers.
As a result, I want to ask whether we need to determine the neuron numbers in hidden layers by our own, and how we can determine it (through experimentation or dice-rolling just as in choosing different models? or by transformer itself?).

[yilmazcemal]

A shared emphasis in these readings, especially in Bommasani et al., seems to concern the "opaqueness" of the foundation models due to their size and complexity where we don't fully understand how the model works or decides. Recently, we also hear news about Chinese AI companies building smaller models with comparable performance to state-of-art LLMs in some part due to constraints in having access to latest available chips. So my question is that whether this might have implications for the transparency of the AI models that we are using, namely can we expect models to be more open to be understandable if their size becomes a point of competition as well and in the long-term starts to decrease, or the decreases in size at those scales do not really matter for model complexity and therefore transparency?

[chychoy]

When discussing activation functions, the Deep Learning textbook discusses how finite ranges for activation functions tend to be more stable, while infinite ranges tend to train more efficiently. What exactly led to this difference, and how are the trade-offs usually balanced in industry settings? Furthermore, will different ranges affect different activation functions differently?

[GitBagguetti]

Looking both the textbook and the Stanford report, I am wondering how we can best use deep learning/foundational models for social scientific research and theory building. I'm wondering about this because the impression I got from the textbook (and partly from the report) is that these models mostly aim to produce a certain decision, output, etc., whereas in the social sciences, we may be more interested in things like interpretation, structure, associations, etc. One way I could see the models' utility is through a study of their internal representations of the training data, but the question is, how is this technically feasible and - given the theoretical difficulty of determining the degree to which foundational models 'understand' the data they are trained on (see Stanford report Ch. 2.6) - done in a theoretically enriching manner?

[giomhern]

I found the concept of layers to be particularly interesting in the nature of neural networks. In the section "What makes a network deep?", the book asserts that "The number of layers determines the depth of the network, and neural networks seemingly perform better with more layers." As my peers also mentioned, interpretability is an issue fundamental to the usefulness of machine learning models as a whole. This issue introduces additional complexities, making me wonder if there is ever or ever will be a model that is too complex. This motivates the following question: Given that neural networks perform better with more generalization via more layers, is there a limit to how deep a model can be? If it can be "infinitely" deep, what are the risks in terms of interpretability or perhaps ethical concerns?

[hchen0628]

In the "Preface" of Thinking with Deep Learning, the authors discuss the "black box" nature of neural networks and their role in generating brain-like representations of data. While they argue that neural networks are not entirely opaque, the text acknowledges that their complexity makes interpretation difficult, especially for multi-layered deep networks.

My question is:
How can social scientists balance the trade-off between leveraging neural networks for their predictive power and the need for interpretability in understanding complex social phenomena? Are there specific techniques (e.g., attention mechanisms, SHAP, LIME) that align better with the goals of social science research, particularly when the focus is on uncovering causal relationships or theoretical insights rather than pure prediction?

[ulisolovieva]

What are the key factors that determine how big (and representative) a dataset needs to be to effectively use DL? For example, could a dataset with 300 data points suffice instead of 500 (like the requirement for HW)? Is it always the case that smaller datasets will need simpler models/less hidden layers? Under what conditions might a smaller dataset still produce reliable results & what strategies could be used to make it work effectively/check viability for DL use?

[xpan4869]

The chapter emphasizes viewing neural networks as "knowledge representations." What advantages does this perspective offer compared to treating them as just predictive models? How can understanding neural networks as knowledge representations help in solving complex, real-world problems with limited data?

[christy133]

I'm really interested in the part of understanding neural networks as knowledge representation. From the first chapter of the book, I learned that neural networks develop internal representations that capture complex patterns, associations, and generalizations through their layered architecture. This process is quite similar to human cognitive processes, such as memory and categorization.

From a societal perspective, as revealed in the inequity and fairness section of the report, these networks develop intrinsic biases embedded within foundation models, which then propagate to multiple downstream applications. This means societal biases present in the training data are amplified during this process.

So, it seems like both human cognition and artificial neural networks naturally develop bias or "cognitive shortcuts" through creating hierarchical representations from environments. Then, what are the implications of models' reliance on biased training data for how they "categorize" and "abstract" knowledge? More specifically, to what extent do these biased representations in AI systems reflect the same cognitive biases present in human thinking, such as stereotyping or overgeneralization? How can we balance between ensuring fairness in these systems and preserving their ability to create meaningful and efficient abstractions in real-life applications?

[xiaotiantangishere]

Should foundation models aim to avoid unintended societal influence? For example, studies suggest that LLMs can affect voters' political leanings through human-LLM interactions. This raises reasonable concerns about the broader, inadvertent societal harm that foundation models might cause. How can we address and mitigate these risks?

[ana-yurt]

As the report on foundation model writes, they often contain "emergent capabilities" (and I guess also emergent defects) that aren't explicitly trained for and can be unexpected. These properties might only reveal when the model is further tuned for more specific tasks. What are some of the practical workflows we could draw on to potentially understand these properties before deploying such models at scale?

[haewonh99]

In the section about Education, it is mentioned that integrating within-domain data(although in various modalities) and outside-domain data(e.g. internet) is useful to make the trained model applicable to various tasks. However, this sounds like incorporating ‘noise’ into the training dataset model built for specific domain expertise in return for flexibility, because the information that is not from the domain may reduce the accuracy of the model in the domain that is intended to be used, or, framed differently, reduce the model’s fit with the domain.

Is this a valid interpretation? That is, does adding this ‘noise’ data from outside the domain reduce model accuracy in the first place? If so, can the inaccuracy caused by the addition of ‘noise’ data be overcome by adding more domain data in parallel to the addition of outside-domain knowledge?

Further, in human learning, the experience of learning about another domain may help with the study of another domain For example, learning to read poetry may actually contribute to the understanding of complicated science theory, as reading literature enhances the ability to comprehend text which is useful in learning in general, including scientific knowledge. Is this not the case for machines?

[Sam-SangJoonPark]

As highlighted in the Stanford report, concerns about data bias remain a critical issue. Predictive models fundamentally rely on patterns from the past to make forecasts about the future. This approach operates on the key assumption that "the future is a repetition of the past." However, if the data is biased, wouldn't it risk reinforcing stereotypes about certain groups, assuming they will always behave in the same way? Could this perpetuate fixed patterns and hinder innovative thinking required for making better decisions? How should we address this issue effectively?

[haonan14]

Foundation models are trained once and deployed widely, often without accounting for new data distributions or societal shifts. The chapter on continual and adaptive learning in the deep learning document presents a provocative idea: could foundation models be designed to evolve over time, learning dynamically from real-world feedback while being audited for ethical performance? This could transform static biases into opportunities for refinement and adaptation.