flâneur

Clementine Yang

1 followers · 913 views

on the atlas — 32

highlights — 219

  • By default, we use the last token representation in this paper.
    2310.01405
  • Thus, a '\n' 'The' 'amount' 'of' 'happiness' 'in' 'the' 'scenario' 'is' ' ' 0 5 10 15 20 25 30 35 Layer Test Accuracy 55 60 65 70 75 80 Figure 5: The representation at the con- cept token “happiness” in middle layers and the representation at the last token in middle and later layers yield high ac- curacy on the utility estimation task. natural position for collecting neural activity associated with concepts is the tokens corresponding to in the task template T c defined in step 1.
    2310.01405
  • The scenarios with high utility evoke distinct neural trajectories within the representation space compared to those with lower utility.
    2310.01405
  • These findings are incompatible with a purely circuit-based account of cognition and are more closely aligned with the Hopfieldian view in cognitive neuroscience
    2310.01405
  • These findings are incompatible with a purely circuit-based account of cognition and are more closely aligned with the Hopfieldian view in cognitive neuroscience
    2310.01405
  • two metrics, usefulness and safety
    2312.06942
  • the red-team instructs GPT-4 to look for good opportunities to insert backdoors and to occasionally take such opportunities when they appea
    2312.06942
  • The blue team (which is standing in for the developers of an AI system) proposes a strategy we call trusted monitoring , where we monitor GPT-4’s outputs with a trusted but less capable model, GPT-3.5.
    2312.06942
  • We want to be robust to intentional subversion from GPT-4 (but we aren’t worried about GPT-3.5)
    2312.06942
  • . Our paper builds on this long line of work by demonstrating that many safety-relevant concepts and processes appear to emerge in LLM representations, enabling us to directly monitor and control these aspects of model cognition via representation engineering.
    2310.01405
  • not limited to text models.
    2310.01405
  • emergent, semantically meaningful internal structure
    2310.01405
  • sgemm_naive >>(M, N, K, alpha, A, B, beta, C);
    How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog
  • For our first kernel, we’ll use the grid, block and thread hierarchy to assign each thread a unique entry in the result matrix C
    How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog
  • or program performance, as we’ll see later, it’s not a good idea to treat all threads in the same block as equals.
    How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog
  • the thread hierarchy we just talked about mostly concerns program correctness
    How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog
  • Threads that are in the same block have access to the same shared memory region (SMEM).
    How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog
  • Each block consists of up to 1024 individual threads
    How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog
  • Each invocation of a CUDA kernel creates a new grid, which consists of multiple blocks
    How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog
  • computation is ordered in a three-level hierarchy
    How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog
  • Note: Assume equal scaling and units for all signals in the plot, if signal scaling or units are not explicitly indicated.
    Scenario trajectory optimization and control on STEP - ScienceDirect
  • A loosely defined precipice for this decision seems to be that for 𝑓 𝐺 𝑊 ≳ 0 . 5 the optimizer will tend toward dominant bootstrap current drive in the high 𝑓 𝐺 𝑊 region and for 𝑓 𝐺 𝑊 ≤ 0 . 5 , the optimizer will choose dominant ECHCD in the low 𝑓 𝐺 𝑊 region
    Scenario trajectory optimization and control on STEP - ScienceDirect
  • where 𝑒 𝑖 is the difference between the plasma current measurement and reference values at time 𝑖 .
    Towards practical reinforcement learning for tokamak magnetic control - ScienceDirect
  • who train mixtures of two policies with an actor-critic approach, learning a single value function for the current mixture of policies. The mixture contains a policy that may be harder to learn but has a higher perfor- mance ceiling
    Growing Action Spaces
  • However, these methods cannot directly be applied to grow the action space in the model-free setting
    Growing Action Spaces
  • Similar curric- ula also appear often in hierarchical reinforcement learning, where skills can be learned in comparatively easy settings and then composed in more complex ways later
    Growing Action Spaces
  • In a self-supervised control setting, Murali et al. (2018) use sensitivity analysis to automatically define a curriculum over action dimensions and prioritise their search space.
    Growing Action Spaces
  • and provides low variance updates
    Growing Action Spaces
  • estimate optimal value functions for multi- ple action spaces simultaneousl
    Growing Action Spaces
  • curriculum of progressively growing action spaces
    Growing Action Spaces
  • During training, the executed control policy is stochastic to explore successful control options. This stochastic policy is represented by a diagonal Gaussian distribution over coil actions.
    Magnetic control of tokamak plasmas through deep reinforcement learning | Nature
  • his was also done in the original DQN paper. This is done mainly to provide more context, i.e. to partially solve the "partial observability" problem
    reinforcement learning - How does one stack multiple observations in the input layer of a convolutional neural network? - Artificial Intelligence Stack Exchange
  • by iteratively updating three components in turn: a policy π (in this case SAC), an encoder φ , and a dynamics model ˆ P
    2006.10742
  • batch Monte Carlo setting for simplicity
    2203.00543
  • provided a theoretical characterisation of how a given representation affects generalization in reinforcement learning
    2203.00543
  • using the same architecture as ours but exchange our bisimulation loss with (1) a reconstruction loss (“ Reconstruction ”) and (2) contrastive predictive coding
    2006.10742
  • we evaluate our method in a clean setting without distractors, as well as a much more difficult setting with distractors. We compare against several baseline
    2006.10742
  • substantially more robust to task-irrelevant distractors
    2006.10742
  • the bisimulation metric is the reward difference added to the Wasserstein distance between transition distributions:
    2006.10742
  • Exact partitioning with bisimulation relations is generally impractical in continuous state spaces, as the relation is highly sensitive to infinitesimal changes in the reward function or dynamics. For this reason, Bisimulation Metrics (Ferns et al., 2011; Ferns & Precup, 2014; Castro, 2020) softens the concept of state partitions, and instead defines a pseudometric space ( S ,d ) , where a distance function d : S×S 7→ R ≥ 0 measures the “behavioral similarity” between two state
    2006.10742
  • two states are bisimilar if they share both the same immediate reward and equivalent distributions over the next bisimilar states
    2006.10742
  • directly learning a representation space with the properties of bisimulation metrics
    2006.10742
  • pproximately equivalent state-action pairs are aggregated
    2006.10742
  • is substantially more robust to complex distractors
    2006.10742
  • by directly learning a bisimilarity-based represen- tation, we guarantee a representation that generates a Lipschitz MDP.
    2006.10742
  • our objective directly learns a representation where distance in latent space is the bisimulation metric
    2006.10742
  • often still require significant manual tuning
    2006.10742
  • jointly train the encoder and the dynamics model end-to-end
    2006.10742
  • n practice, there is no guarantee that the learned representation will capture useful information for the control task,
    2006.10742
  • when compared to prior approaches that use reconstruction losses or contrastive losses
    2006.10742