Skip to content
TheChatArchiveThe Chat Archive
o4-mini-2025-04-16· OpenAIHugging Face LMArena Human Preference 140K

Design a computational graph for "Recurrent networks with recurrent connections between…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with DeepSeekdeepseek-r1-0528’s answer.
  1. User

    Design a computational graph for "Recurrent networks with recurrent connections between hidden units, that read an entire sequence and then produce a single output" and develop the forward propagation equations for the same.

  2. Assistant

    Here’s a simple unrolled computational graph for an RNN that reads a sequence
    ({x_1,\dots,x_T}) and—after the last time step—produces one output (y).

    flowchart LR
      direction LR
      subgraph Unrolled_RNN
        h0(["h₀"])
        x1(["x₁"])
        h1(["h₁"])
        x2(["x₂"])
        h2(["h₂"])
        x3(["x₃"])
        h3(["h₃"])
        XT(["..."])
        xT(["x_T"])
        hT(["h_T"])
        y (["y"])
      
        h0 --> h1
        x1 --> h1
        h1 --> h2
        x2 --> h2
        h2 --> h3
        x3 --> h3
        h3 --> XT
        XT --> xT
        xT --> hT
        h(T-1)(["h_{T-1}"]) --> hT
        hT --> y
      end
    

    Legend:

    • (x_t\in\R^{d_x}) is the input at time (t).
    • (h_t\in\R^{d_h}) is the hidden state after seeing (x_t).
    • (y\in\R^{d_y}) is the final output (only after time (T)).

    Forward-Propagation Equations

    1. Initialization
      [ h_0 = \mathbf{0}\quad(\text{often a zero‐vector}) ]

    2. Recurrence (for (t=1,\dots,T))
      [ h_t = \sigma_h!\bigl(W_{xh},x_t ;+; W_{hh},h_{t-1} ;+; b_h\bigr) ] • (W_{xh}\in\R^{d_h\times d_x})
      • (W_{hh}\in\R^{d_h\times d_h})
      • (b_h\in\R^{d_h})
      • (\sigma_h(\cdot)) = hidden activation (e.g.\ (\tanh) or ReLU)

    3. Output (after (t=T))
      [ y = \sigma_y!\bigl(W_{hy},h_T ;+; b_y\bigr) ] • (W_{hy}\in\R^{d_y\times d_h})
      • (b_y\in\R^{d_y})
      • (\sigma_y(\cdot)) = output activation
      (e.g.\ softmax for classification, identity for regression)


    Walk-Through (toy example, (T=3))

    1. (h_1 = \sigma_h(W_{xh}x_1 + W_{hh}h_0 + b_h))
    2. (h_2 = \sigma_h(W_{xh}x_2 + W_{hh}h_1 + b_h))
    3. (h_3 = \sigma_h(W_{xh}x_3 + W_{hh}h_2 + b_h))
    4. (y = \sigma_y(W_{hy}h_3 + b_y))

    That’s it—by unrolling the recurrence you get a directed acyclic “computational graph,”
    and the above equations perform one forward pass from ({x_t}) → ({h_t}) → (y).

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.