Sumeet Seth

Epidemiology and Machine learning are not always in consonance. When looking at data for deductive interpretations, it is very important to consider causality. ML research is not very worried about causation, but epidemiology is quite the opposite. The following article by Blakely et al. brings out how they could align, but the real world applications are not necessarily being so rigorous. Be very careful! [Blakely, T., Lynch, J., Simons, K., Bentley, R., & Rose, S. (2019). Reflection on Modern Methods: When Worlds Collide—Prediction, Machine Learning and Causal Inference. International Journal of Epidemiology, 1–7. https://doi.org/10.1093/ije/dyz132]

I will talk about the two areas I believe are the most important - A) framing a problem using story-telling, to capture the imagination of your audience. Numbers hold a story that you have to bring out. Use pictures, tools or anecdotes. B) Focus on innovation - do many small things in better ways when looking at an existing mature space, if you can't get a big idea that will change how people look at it completely.

In medicine, events can be things like a patient having a particular disease (D), or a test result being positive (T). Conditional Probability is the probability of one event (let's say the test being positive) happening given that another event (the patient having the disease) has already happened. It's written as P(T|D). Imagine a disease with a 1% prevalence in the population (P(D) = 0.01). A new test is developed to detect. The test is very good at catching the disease (high sensitivity, P(T|D) = 0.9), but it can give a positive result for people who don't have the disease (low specificity, P(T|~D) = 0.1). Bayes' theorem can calculate the positive predictive value (PPV), probability the patient has the disease given a positive test.

While there are many aspects of how to choose a model, the following approach can be used as a starting point - Define your goal - Are you forecasting sales (regression) or identifying fraudulent transactions (classification)? Examine your data - Is it mostly linear relationships, or are there complex patterns? Is there enough data to train complex models? Consider interpretability - If you need to explain the model's reasoning, linear regression or decision trees might be better choices. Think about resources - Deep learning models can be powerful but require significant computing power.

Here's a simple comparison of linear regression and other predictive models: Relationship - Linear regression assumes a straight line relationship between the independent and dependent variables. Other models can capture more complex relationships, like curves or exponential growth. Output - Linear regression predicts continuous values. Other models might predict categories (logistic regression) or even entirely new data points (generative models). Interpretability - Linear regression is often easier to interpret because the coefficients directly reflect the impact of each variable. Other models can be more like black boxes. Data Complexity - Linear regression works well for simpler data patterns.

Deep learning in design tools is a double-edged sword. While it offers exciting possibilities, challenges remain. Gathering massive, clean data for training is expensive and can lead to bias. The "black box" nature of deep learning models makes it hard for designers to understand and control the AI's suggestions. There's also the need for powerful computers to run these models, and the field is constantly evolving, requiring ongoing updates to design tools. Despite these hurdles, the potential for deep learning to transform design is significant.