UEMA trains networks to connect what its body does with what its senses report next. The resulting patterns can become useful distinctions: contact, resistance, and the effects of its own movements. Other networks learn its bodily condition and predict how that condition will change. This gives our consciousness thesis a specific learning mechanism.
MIMo is the simulated body. UEMA is the learning system. A reach supplies several kinds of information about the same event:
Keep the senses, their times and the executed command together.
MIMo bridgeMultiply and combine readings using learned weights. Produce a compact list of numbers, called a token.
EncoderRegistryAttention combines tokens across senses and time. Stacked neural blocks build on earlier combinations.
ForwardModel / ThreeStageBlockEncode what really happened. Hold these target numbers fixed for this lesson.
For each value: subtract the actual token from the predicted token, square the difference, then average.
Small networks produce the predicted next token for each predicted sense.
PredictionHeadsWhat creates the new perception? A difference matters to this lesson when preserving it helps predict a different outcome. Training can therefore build a useful distinction such as contact versus no contact without supplying those words as labels. Later network layers can use that distinction to represent more complex relationships. This is function begets function.
A perception space contains every observation the senses can represent. Each observation is one point in that space.
The tokens can stay the same size while learning changes what differences they preserve.
The touch encoder takes a row of numbers for each skin site. A small network multiplies these numbers by learned weights, adds biases and applies a curved response called GELU. Further weighted layers combine the results. Attention then mixes information between sites. A learned pooling query combines the site results into one touch token. Where configured, a learned site identifier tells the encoder which skin location supplied a row.
Source: CrossPatchAttentionEncoder, patch_mlp, layers and pool_attn. The prediction lesson changes these trainable weights, so the retained touch patterns can change with experience.
Attention computes learned matching scores between tokens, turns those scores into weights, and mixes their information. Each forward-model block has three attention stages. Select a stage to see which tokens it can combine.
Blue: the token being updated. Green: tokens it can read in this stage. Four input streams are shown. Colors show allowed connections, not measured attention strength.
The master configuration repeats these stages eight times. Each block also has a small layered network that transforms the combined values and adds its output to the incoming values. Block count is configurable; meanings such as “grasp” are learned patterns, not assigned block names.
Prediction error trains the projections. The diagram shows which connections are allowed; training determines their weights.
The time-only stage blocks later times. The across-both stage can read the entire supplied window. For the next-reading lesson shown here, that window ends before the reading being predicted. The answer enters only when the error is calculated. A lesson predicting every point inside a longer recording needs past-only windows or a time mask on every path that mixes times.
UEMA tracks which changes accompany its commands and learns what condition its body is in. Wellness describes functioning such as sensing quality and available motor capacity. The two paths below connect body experience to that evaluation.
This path uses running statistics. An outside object's membership also compares its reported movement with body movement. The extent route uses body_schema_updates=on and wellness_coherence=extent.
perceive mode →Main model's latest tokensThe added values reach prediction heads and movement computations. learn mode teaches self state without adding it to the main model. A recurrent network carries state forward from one moment to the next.
With self_prediction=grounded_residual, the forecast starts from measured current bodily function. The separate learned_present output is still trained against measurements. The present measurement, the learned self representation and the future forecast are three different things.
PersistentMemory combines stored events with their elapsed times using a recurrent network, then uses learned attention to retrieve prior states. A learned readout adds information to the forward model's final tokens. This is a separate memory path from the self network's own retained state.
The prediction system can represent consequences; the self mechanisms connect consequences to this body's control and condition. Intention uses learned predictions to search for commands whose consequences match a goal. Wellness helps select which remembered outcomes become goals.
This is the body_function=modulate path. Recovery can restore precision. In the same mode, fatigue reduces available motor force, changing the movement and the next sensed outcome.
Body state can change what is sensed, its learned representation and how learning proceeds. Separate optional activity settings change sensory gain and learning rate. Gain scales values; it is different from the touch-precision path above. The broader patent design also covers changes to computing resources.
In Many Minds, a specialist receives selected sensory inputs while the parent retains the full set. Mastered skills retain prediction and wellness within their own minds. This extends the learned self across a hierarchy of abilities.
Our thesis is that conscious awareness can emerge from this developing, self-related perception. The model learns distinctions about the world, learns its own place and condition within that world, and uses those representations to shape further perception and action. These connected learning processes are the substance of the proposal.
| Connection | Source inspected for this revision |
|---|---|
| Sensor readings to learned tokens | EncoderRegistry; touch/motor encoder |
| Three attention stages, layered transforms | ForwardModel / ThreeStageBlock; master configuration |
| Next-reading targets and weight updates | MIMo model, next-frame lesson; prediction heads |
| Body membership and learned self state | body schema; SelfState; self integration |
| Memory and intention | persistent memory; intention generator |
| Changed sensing, force and learning settings | body function; environment; activity modulation; wellness |
Thesis: Perception Space v4. Wider architecture: UEMA specification v3.6. Modes beside each route identify when it contributes.
Reviewed November 1, 2026 by Codex and two independent reviewers. Source and clarity checks corrected encoder detail, self-feedback arrows, body-membership wording and the phone comparison layout. Desktop and phone views inspected; attention controls checked.