| . | Held-out t = 1 . | Held-out t = 2 . | Held-out t = 3 . | Held-out t = 4 . | ||||
|---|---|---|---|---|---|---|---|---|
| Model . | Train . | Test . | Train . | Test . | Train . | Test . | Train . | Test . |
| TrajectoryNet | 10.59 ± 1.08 | 12.81 ± 0.08 | 10.71 ± 1.01 | 11.69 ± 0.24 | 11.51 ± 0.49 | 9.62 ± 0.14 | 11.09 ± 0.83 | 10.38 ± 0.13 |
| MIOFlow | 10.09 ± 0.50 | 10.91 ± 0.02 | 10.33 ± 0.47 | 10.98 ± 0.22 | 10.37 ± 0.35 | 9.31 ± 0.18 | 10.34 ± 0.51 | 10.31 ± 0.11 |
| PRESCIENT(+g)a | 10.18 ± 1.19 | 11.45 ± 0.04 | 9.86 ± 1.16 | 9.54 ± 0.09 | 10.46 ± 1.10 | 8.11 ± 0.06 | 9.93 ± 1.08 | 9.06 ± 0.09 |
| PRESCIENT | 7.83 ± 0.37 | 10.45 ± 0.11 | 7.96 ± 0.37 | 8.91 ± 0.02 | 8.22 ± 0.37 | 7.52 ± 0.08 | 8.17 ± 0.34 | 7.79 ± 0.09 |
| PI-SDE | 7.36 ± 0.32∗ | 10.36 ± 0.05 | 7.36 ± 0.40 | 8.35 ± 0.12 | 7.65 ± 0.39 | 7.41 ± 0.03 | 7.69 ± 0.44 | 7.61 ± 0.35 |
| . | Held-out t = 1 . | Held-out t = 2 . | Held-out t = 3 . | Held-out t = 4 . | ||||
|---|---|---|---|---|---|---|---|---|
| Model . | Train . | Test . | Train . | Test . | Train . | Test . | Train . | Test . |
| TrajectoryNet | 10.59 ± 1.08 | 12.81 ± 0.08 | 10.71 ± 1.01 | 11.69 ± 0.24 | 11.51 ± 0.49 | 9.62 ± 0.14 | 11.09 ± 0.83 | 10.38 ± 0.13 |
| MIOFlow | 10.09 ± 0.50 | 10.91 ± 0.02 | 10.33 ± 0.47 | 10.98 ± 0.22 | 10.37 ± 0.35 | 9.31 ± 0.18 | 10.34 ± 0.51 | 10.31 ± 0.11 |
| PRESCIENT(+g)a | 10.18 ± 1.19 | 11.45 ± 0.04 | 9.86 ± 1.16 | 9.54 ± 0.09 | 10.46 ± 1.10 | 8.11 ± 0.06 | 9.93 ± 1.08 | 9.06 ± 0.09 |
| PRESCIENT | 7.83 ± 0.37 | 10.45 ± 0.11 | 7.96 ± 0.37 | 8.91 ± 0.02 | 8.22 ± 0.37 | 7.52 ± 0.08 | 8.17 ± 0.34 | 7.79 ± 0.09 |
| PI-SDE | 7.36 ± 0.32∗ | 10.36 ± 0.05 | 7.36 ± 0.40 | 8.35 ± 0.12 | 7.65 ± 0.39 | 7.41 ± 0.03 | 7.69 ± 0.44 | 7.61 ± 0.35 |
The table presents results from four distinct held-out tasks, where Day 1, Day 2, Day 3, and Day 4 were excluded from the training process, respectively. For each task, we compute the average Wasserstein distance between observed data and predicted data (training loss) and Wasserstein distance between unseen data and predicted data (test loss).
The bold values imply the best performance.
PRESCIENT with estimated growth rate.
| . | Held-out t = 1 . | Held-out t = 2 . | Held-out t = 3 . | Held-out t = 4 . | ||||
|---|---|---|---|---|---|---|---|---|
| Model . | Train . | Test . | Train . | Test . | Train . | Test . | Train . | Test . |
| TrajectoryNet | 10.59 ± 1.08 | 12.81 ± 0.08 | 10.71 ± 1.01 | 11.69 ± 0.24 | 11.51 ± 0.49 | 9.62 ± 0.14 | 11.09 ± 0.83 | 10.38 ± 0.13 |
| MIOFlow | 10.09 ± 0.50 | 10.91 ± 0.02 | 10.33 ± 0.47 | 10.98 ± 0.22 | 10.37 ± 0.35 | 9.31 ± 0.18 | 10.34 ± 0.51 | 10.31 ± 0.11 |
| PRESCIENT(+g)a | 10.18 ± 1.19 | 11.45 ± 0.04 | 9.86 ± 1.16 | 9.54 ± 0.09 | 10.46 ± 1.10 | 8.11 ± 0.06 | 9.93 ± 1.08 | 9.06 ± 0.09 |
| PRESCIENT | 7.83 ± 0.37 | 10.45 ± 0.11 | 7.96 ± 0.37 | 8.91 ± 0.02 | 8.22 ± 0.37 | 7.52 ± 0.08 | 8.17 ± 0.34 | 7.79 ± 0.09 |
| PI-SDE | 7.36 ± 0.32∗ | 10.36 ± 0.05 | 7.36 ± 0.40 | 8.35 ± 0.12 | 7.65 ± 0.39 | 7.41 ± 0.03 | 7.69 ± 0.44 | 7.61 ± 0.35 |
| . | Held-out t = 1 . | Held-out t = 2 . | Held-out t = 3 . | Held-out t = 4 . | ||||
|---|---|---|---|---|---|---|---|---|
| Model . | Train . | Test . | Train . | Test . | Train . | Test . | Train . | Test . |
| TrajectoryNet | 10.59 ± 1.08 | 12.81 ± 0.08 | 10.71 ± 1.01 | 11.69 ± 0.24 | 11.51 ± 0.49 | 9.62 ± 0.14 | 11.09 ± 0.83 | 10.38 ± 0.13 |
| MIOFlow | 10.09 ± 0.50 | 10.91 ± 0.02 | 10.33 ± 0.47 | 10.98 ± 0.22 | 10.37 ± 0.35 | 9.31 ± 0.18 | 10.34 ± 0.51 | 10.31 ± 0.11 |
| PRESCIENT(+g)a | 10.18 ± 1.19 | 11.45 ± 0.04 | 9.86 ± 1.16 | 9.54 ± 0.09 | 10.46 ± 1.10 | 8.11 ± 0.06 | 9.93 ± 1.08 | 9.06 ± 0.09 |
| PRESCIENT | 7.83 ± 0.37 | 10.45 ± 0.11 | 7.96 ± 0.37 | 8.91 ± 0.02 | 8.22 ± 0.37 | 7.52 ± 0.08 | 8.17 ± 0.34 | 7.79 ± 0.09 |
| PI-SDE | 7.36 ± 0.32∗ | 10.36 ± 0.05 | 7.36 ± 0.40 | 8.35 ± 0.12 | 7.65 ± 0.39 | 7.41 ± 0.03 | 7.69 ± 0.44 | 7.61 ± 0.35 |
The table presents results from four distinct held-out tasks, where Day 1, Day 2, Day 3, and Day 4 were excluded from the training process, respectively. For each task, we compute the average Wasserstein distance between observed data and predicted data (training loss) and Wasserstein distance between unseen data and predicted data (test loss).
The bold values imply the best performance.
PRESCIENT with estimated growth rate.
This PDF is available to Subscribers Only
View Article Abstract & Purchase OptionsFor full access to this pdf, sign in to an existing account, or purchase an annual subscription.
