Berker Demirel

Research

I study the geometry and structure of neural representations: how they are organized, when they collapse, and how they behave under distribution shift. My research focuses on three closely related areas. Representation learning looks at what information representations contain, how easily it can be read out and how to learn useful ones. Generalization is about how well models perform outside their training distribution and how to detect when they fail. Generative modeling focuses on learning data distributions, generating new samples and evaluating their quality.

These areas are closely connected, as the quality and structure of learned representations affect how well models generalize, how we learn data distributions and how we evaluate generative models.

01Representation learning

A representation turns data into features that other models can use. The challenge is that we often don't know ahead of time what those models will need to do. Ideally, a good representation should be useful across many different tasks, while also making the relevant information easy to extract, even with something as simple as a linear probe.

Representation

  • circle
  • triangle
  • color 1
  • color 2
  • linear probe

Probe accuracy

  • linear probe
  • flexible probe (kNN)

Open questions

  • What makes a representation good if we don't know what it will eventually be used for?
  • How can we learn representations that are both broadly useful and easy to work with, ideally without labels?
  • How does the way information is spread across different directions affect what we can actually read from a representation?

Our work

02Generalization

Generalization is about how well a model performs on data outside its training distribution. Since training data can never cover every possible situation, generalization depends on what changes between training and testing, what stays the same, and which changes can be identified from the inputs. When a model fails to generalize, we also want to be able to detect that, including across different learning paradigms.

Inputs

  • training data
  • test data
  • class A
  • class B
  • learned boundary
  • true boundary
  • misclassified

As the shift grows

  • accuracy
  • confidence
  • flagged OOD
accuracy – confidence – flagged OOD –
inputs
–
labeling rule
–

Open questions

  • When can we expect a model to generalize, and what does that depend on?
  • Can we tell when a model fails to generalize?
  • How do different types of models, such as classifiers and generative models, respond to the same changes in the data?

Our work

03Generation

Generative modeling is about learning a data distribution so that we can generate new samples from it. Since a model only sees a limited amount of training data, the challenge is to learn the underlying distribution rather than simply reproduce what it has seen. Evaluating the results is not straightforward either. With high-dimensional data, there are many ways a model can get the distribution wrong and limited real data makes these problems hard to detect. A good score on one metric may therefore hide important problems. This is why we need to assess the quality of generated samples in different ways.

Sampling

  • training point
  • sampling trajectory
  • generated sample
  • near-copy of a training point

Evaluation

Open questions

  • How can we learn a data distribution? Approaches like diffusion models, latent diffusion models, GANs and drifting models do this in different ways.
  • Can a model generate genuinely new samples or does it mostly memorize its training data?
  • How can we tell if generated samples are good? Distributional metrics, classifier-based evaluation, qualitative assessment and independent domain-specific evaluators each test different aspects of quality.

Our work