ABSTRACT We explore the use of deep learning to infer the temperature of the intergalactic medium from the transmitted flux in the high-redshift Ly α forest. We train neural networks on sets of simulated spectra from redshift z = 2–3 outputs of cosmological hydrodynamic simulations, including high-temperature regions added in post-processing to approximate bubbles heated by He ii reionization. We evaluate how well the trained networks are able to reconstruct the temperature from the effect of Doppler broadening in the simulated input Ly α forest absorption spectra. We find that for spectra with high resolution (10 $\, {\rm km}\, {\rm s}^{-1}$ pixel) and moderate signal-to-noise ratio (20–50), the neural network is able to reconstruct the intergalactic medium temperature smoothed on scales of $\sim 6 \, h^{-1}\, {\rm Mpc}$ quite well. Concentrating on discontinuities, we find that high-temperature regions of width $25 \, h^{-1}\, {\rm Mpc}$ and temperature $20\, 000$ K can be fairly easily detected and characterized. We show an example where multiple sightlines are combined to yield tomographic images of hot bubbles. Deep learning techniques may be useful in this way to help us understand the complex temperature structure of the intergalactic medium around the time of helium reionization.