Individual differences among deep neural network models

By Johannes Mehrer, Courtney J. Spoerer, Nikolaus Kriegeskorte, Tim C Kietzmann. 2020

181 citations recorded by the source corpus.

Deep neural networks (DNNs) excel at visual recognition tasks and are increasingly used as a modeling framework for neural computations in the primate brain. Just like individual brains, each DNN has a unique connectivity and representational profile. Here, we investigate individual differences among DNN instances that arise from varying only the random initialization of the network weights. Using tools typically employed in systems neuroscience, we show that this minimal change in initial conditions prior to training leads to substantial differences in intermediate and higher-level network representations despite similar network-level classification performance. We locate the origins of the effects in an under-constrained alignment of category exemplars, rather than misaligned category centroids. These results call into question the common practice of using single networks to derive insights into neural information processing and rather suggest that computational neuroscientists working with DNNs may need to base their inferences on groups of multiple network instances. Do artificial neural networks, like brains, exhibit individual differences? Using tools from systems neuroscience, this study reveals substantial variability in network-internal representations, calling into question the neuroscientific practice of using single networks as models of brain function.

Subfields: Cognitive Neuroscience

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Cognitive Neuroscience

Citations in the corpus

Orthogonal representations for robust context-dependent task performance in brains and neural networks (2022, positive): “More generally, the diverse representational structure that can emerge in the rich and lazy regimes, and its variablemapping to the brain, may shed light on why emerging representation structure can be heterogenous in trained neural networks (Mehrer et al., 2020).”
Beyond accuracy: quantifying trial-by-trial behaviour of CNNs and humans by measuring error consistency (2020, unknown): “It remains an open question why even multiple instances of a single model (trained with a different random seed) internally often differ substantially [50, 51], yet in spite of large architectural differences across models and model families, all CNNs that we investigated seem to be implementing fai”
Model Multiplicity: Opportunities, Concerns, and Solutions (2022, background): “And yet, empirical evidence seems to suggest the opposite: developing more accurate models can often lead tomore multiplicity (see Figure 1 for an example) [10, 41].”
What shapes feature representations? Exploring datasets, architectures, and training (2020, unknown): “tures a model learns, and how consistent feature representations are across model instances, is relevant to understanding how models compute [2], as well as how they should be compared to one another [29] or to neural data [19, 6, 5, 39, 43]. In this paper, we investigate the evolution of models’ fe”
Why Do Better Loss Functions Lead to Less Transferable Features? (2020, background): “[40] previously reported increased class separation in CIFAR-10 networks trained with dropout.”