Dictionary
Concept dictionary
The vocabulary you need to read this archive. Each entry is the term as the field uses it, paired with a paper in the corpus where Lerchner works with it.
The definitions are ours, written to get you into the papers — they are not quotations of Lerchner. The citation under each one is the part that is his.
- Representation
- What a system carries about the world internally — the quantity every phase of this archive is about, whether it lives in a cortical population, a latent vector, an object slot, or an agent’s policy.
- In the corpus Towards a Definition of Disentangled Representations (2018) ↗
- Disentangled representation
- A representation in which separate, independent factors of variation in the data occupy separate parts of the code — so changing one dimension changes one thing about the reconstruction. Widely used before it was defined; this corpus contains the paper that attempts the definition.
- In the corpus Towards a Definition of Disentangled Representations (2018) ↗
- Variational autoencoder (VAE)
- A generative model that learns to compress data into a probability distribution over latent variables and reconstruct it, trained by maximizing a bound on the data likelihood (the ELBO). The substrate for the disentangling work here.
- In the corpus beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework (2017) ↗
- β (beta)
- The single added coefficient that gives β-VAE its name: it weights the latent bottleneck term against reconstruction. Push it above one and the representation trades reconstruction detail for independent, interpretable factors — the whole result in one hyperparameter.
- In the corpus Understanding disentangling in $\beta$-VAE (2018) ↗
- Factors of variation
- The independent generative causes behind a dataset — position, colour, scale, identity — that a disentangled model is supposed to recover without ever being told they exist.
- In the corpus Early Visual Concept Learning with Unsupervised Deep Learning (2016) ↗
- Compositionality
- The property that learned parts can be recombined into configurations never seen in training. SCAN pursues it hierarchically, on the conjecture that the world’s diversity "arises from a relatively small set of coherent rules".
- In the corpus SCAN: Learning Hierarchical Compositional Visual Concepts (2018) ↗
- Zero-shot transfer
- Performing in a domain the agent was never trained on, with no additional learning. DARLA’s claim is that a disentangled visual representation is what makes it possible — the agent "learns to see before learning to act".
- In the corpus DARLA: Improving Zero-Shot Transfer in Reinforcement Learning (2017) ↗
- Object-centric representation
- Representing a scene as a set of discrete entities with their own properties, rather than as one undifferentiated feature map — argued here to be "crucial for general intelligence" because it is what lets reasoning and imagination recombine parts.
- In the corpus MONet: Unsupervised Scene Decomposition and Representation (2019) ↗
- Scene decomposition
- Splitting an observation into its constituent objects without supervision — no segmentation labels, no object list. MONet does it per image; SIMONe does it over video while simultaneously inferring the viewpoint.
- In the corpus SIMONe: View-Invariant, Temporally-Abstracted Object Representations via Unsupervised Video Decomposition (2021) ↗
- Meta-reinforcement learning
- Learning to learn: training an agent so that it can adapt quickly to a new task drawn from a family, rather than mastering one fixed task. Alchemy exists because the field lacked benchmarks whose underlying structure was neither too simple nor too ill-defined.
- In the corpus Alchemy: A benchmark and analysis toolkit for meta-reinforcement learning agents (2021) ↗
- Instructable agent
- An agent that follows free-form natural-language instructions in an environment, rather than optimizing a fixed reward. The SIMA project trains such agents across many commercial 3D worlds.
- In the corpus Scaling Instructable Agents Across Many Simulated Worlds (2024) ↗
- Balanced cortical network
- A model of cortex in which strong excitation and strong inhibition roughly cancel, leaving neurons in a fluctuation-driven, irregular firing regime. The earliest work in this archive develops mean-field theories that solve such networks for firing rates and correlations together.
- In the corpus Response Variability in Balanced Cortical Networks (2006) ↗