A hologram can make an image. A recent hybrid optical experiment instead uses a light pattern to identify image content: a bright region represents a selected answer, and separate regions serve different tasks.[2]
The work, published online on 25 September in Opto-Electronic Technology and highlighted in an October research release, uses a single laser, a phase-only spatial light modulator and a camera. Its computing system is hybrid. An electronic network prepares the pattern imposed on the light; the light’s propagation produces the distribution read at the other end.[1][2]
The answer has an address
Imagine two answer boards: one for digits, another for clothing categories. The task is to direct light to the appropriate position on each. The camera records intensity, not a sentence explaining the prediction.[1]
The publisher’s account describes joint electronic and optical training rather than physically rearranging the apparatus for each task. The electronic-to-optical conversion is part of the learned computation. Calling the system a holographic neural network should not suggest that an unprocessed photograph simply shines through a lens and becomes understood.[2]
In conceptual terms, this changes what is being optimised. A conventional screen is judged by whether its visible image resembles a target picture. A classifier’s output can instead be judged by whether the light reaches the region assigned to the correct category. The optics need not produce a recognisable picture to produce a useful signal.
What the numerical comparison actually shows
Simulated blind tests of the same trained network
| Test input | Digit accuracy | Fashion accuracy |
|---|---|---|
| Digit plus fashion image in the trained arrangement | 91.35% | 83.16%Source 1 |
| Digit only | 91.99% | Not applicableSource 1 |
| Digit plus relocated or resized fashion image | 91.00% | Not reported in this tableSource 1 |
The comparison asks a specific question: how much does disturbing one part of an input affect the other task? The digit score stays close to its original value. That is different from demonstrating that both tasks handle arbitrary changes equally well; a stable result in one output does not supply the missing result for the other.
The paper also presents physical output examples. Its spatial light modulator refreshes at up to 60 Hz, and it discusses camera integration and calibration. The fact that light propagates quickly therefore does not establish the speed or energy efficiency of the complete electronic–optical–electronic process.[1]
A different way to divide the work
The attraction is architectural: a learned electronic front end can make a comparatively simple optical path serve several outputs. The experiment does not need to compete with every general-purpose AI model to make that idea worth exploring. It asks where a physical transformation can do a useful part of a computation, and what information must be prepared before it arrives there.
The hologram in this story is not the final picture. It is the arrangement that helps the light deliver two answers to two different places.
Sources & further reading
- Neural holography enabled multi-beam steering for multi-task classifier Wu and colleagues / Opto-Electronic Technology · 25 Sep 2026 · Accessed 14 Oct 2026
- New holographic neural network uses light to perform multiple AI tasks at once Opto-Electronic Technology / EurekAlert! · 13 Oct 2026 · Accessed 14 Oct 2026